diff options
| -rw-r--r-- | perform/pio_engine.c | 60 | ||||
| -rw-r--r-- | perform/pio_perf.c | 163 | ||||
| -rw-r--r-- | perform/pio_perf.h | 9 | ||||
| -rw-r--r-- | perform/pio_timer.c | 43 | ||||
| -rw-r--r-- | perform/pio_timer.h | 2 |
5 files changed, 141 insertions, 136 deletions
diff --git a/perform/pio_engine.c b/perform/pio_engine.c index 79f035a..0c5a9f3 100644 --- a/perform/pio_engine.c +++ b/perform/pio_engine.c @@ -145,7 +145,7 @@ static void do_cleanupfile(iotype iot, char *fname); * Modifications: */ results -do_pio(FILE *output, parameters param) +do_pio(parameters param) { /* return codes */ herr_t ret_code = 0; /*return code */ @@ -164,10 +164,6 @@ do_pio(FILE *output, parameters param) /* HDF5 variables */ herr_t hrc; /*HDF5 return code */ - int myrank; - - MPI_Comm_rank(pio_comm_g, &myrank); - /* Sanity check parameters */ /* IO type */ @@ -256,10 +252,15 @@ buf_size=MIN(1024*1024, buf_size); GOTOERROR(FAIL); } - if (pio_debug_level >= 4) + if (pio_debug_level >= 4) { + int myrank; + + MPI_Comm_rank(pio_comm_g, &myrank); + /* output all of the times for all iterations */ if (myrank == 0) fprintf(output, "Timer details:\n"); + } for (nf = 1; nf <= nfiles; nf++) { /* @@ -277,43 +278,20 @@ fprintf(stderr, "filename=%s\n", fname); #endif set_time(res.timers, HDF5_GROSS_WRITE_FIXED_DIMS, START); - - if (pio_debug_level >= 4) - /* output all of the times for all iterations */ - fprintf(output, " Proc %d: Gross Write Start: %.2f\n", - myrank, get_time(res.timers, HDF5_GROSS_WRITE_FIXED_DIMS)); - hrc = do_fopen(iot, fname, &fd, PIO_CREATE | PIO_WRITE); VRFY((hrc == SUCCESS), "do_fopen failed"); set_time(res.timers, HDF5_FINE_WRITE_FIXED_DIMS, START); - - if (pio_debug_level >= 4) - /* output all of the times for all iterations */ - fprintf(output, " Proc %d: Fine Write Start: %.2f\n", - myrank, get_time(res.timers, HDF5_FINE_WRITE_FIXED_DIMS)); - hrc = do_write(&fd, iot, ndsets, nelmts, buf_size, buffer); set_time(res.timers, HDF5_FINE_WRITE_FIXED_DIMS, STOP); - if (pio_debug_level >= 4) - /* output all of the times for all iterations */ - fprintf(output, " Proc %d: Fine Write Stop: %.2f\n", - myrank, get_time(res.timers, HDF5_FINE_WRITE_FIXED_DIMS)); - VRFY((hrc == SUCCESS), "do_write failed"); /* Close file for write */ hrc = do_fclose(iot, &fd); set_time(res.timers, HDF5_GROSS_WRITE_FIXED_DIMS, STOP); - - if (pio_debug_level >= 4) - /* output all of the times for all iterations */ - fprintf(output, " Proc %d: Gross Write Stop: %.2f\n", - myrank, get_time(res.timers, HDF5_GROSS_WRITE_FIXED_DIMS)); - VRFY((hrc == SUCCESS), "do_fclose failed"); MPI_Barrier(pio_comm_g); @@ -323,43 +301,19 @@ fprintf(stderr, "filename=%s\n", fname); */ /* Open file for read */ set_time(res.timers, HDF5_GROSS_READ_FIXED_DIMS, START); - - if (pio_debug_level >= 4) - /* output all of the times for all iterations */ - fprintf(output, " Proc %d: Gross Read Start: %.2f\n", - myrank, get_time(res.timers, HDF5_GROSS_READ_FIXED_DIMS)); - hrc = do_fopen(iot, fname, &fd, PIO_READ); VRFY((hrc == SUCCESS), "do_fopen failed"); set_time(res.timers, HDF5_FINE_READ_FIXED_DIMS, START); - - if (pio_debug_level >= 4) - /* output all of the times for all iterations */ - fprintf(output, " Proc %d: Fine Read Start: %.2f\n", - myrank, get_time(res.timers, HDF5_FINE_READ_FIXED_DIMS)); - hrc = do_read(&fd, iot, ndsets, nelmts, buf_size, buffer); set_time(res.timers, HDF5_FINE_READ_FIXED_DIMS, STOP); - - if (pio_debug_level >= 4) - /* output all of the times for all iterations */ - fprintf(output, " Proc %d: Fine Read Stop: %.2f\n", - myrank, get_time(res.timers, HDF5_FINE_READ_FIXED_DIMS)); - VRFY((hrc == SUCCESS), "do_read failed"); /* Close file for read */ hrc = do_fclose(iot, &fd); set_time(res.timers, HDF5_GROSS_READ_FIXED_DIMS, STOP); - - if (pio_debug_level >= 4) - /* output all of the times for all iterations */ - fprintf(output, " Proc %d: Gross Read Stop: %.2f\n", - myrank, get_time(res.timers, HDF5_GROSS_READ_FIXED_DIMS)); - VRFY((hrc == SUCCESS), "do_fclose failed"); MPI_Barrier(pio_comm_g); diff --git a/perform/pio_perf.c b/perform/pio_perf.c index db6ccc0..7dca0e4 100644 --- a/perform/pio_perf.c +++ b/perform/pio_perf.c @@ -79,8 +79,10 @@ #define MB_PER_SEC(bytes,t) (((bytes) / ONE_MB) / t) /* global variables */ -int comm_world_rank_g; /* my rank in MPI_COMM_RANK */ -int comm_world_nprocs_g;/* num. of processes of MPI_COMM_WORLD */ +FILE *output; /* output file */ +pio_time *timer; /* timer: global for stub functions */ +int comm_world_rank_g; /* my rank in MPI_COMM_RANK */ +int comm_world_nprocs_g;/* num. of processes of MPI_COMM_WORLD */ MPI_Comm pio_comm_g; /* Communicator to run the PIO */ int pio_mpi_rank_g; /* MPI rank of pio_comm_g */ int pio_mpi_nprocs_g; /* Number of processes of pio_comm_g */ @@ -222,15 +224,15 @@ typedef struct _minmax { /* local functions */ static long parse_size_directive(const char *size); static struct options *parse_command_line(int argc, char *argv[]); -static void run_test_loop(FILE *output, struct options *options); -static int run_test(FILE *output, iotype iot, parameters parms); -static void output_all_info(FILE *output, minmax *mm, int count, int indent_level); +static void run_test_loop(struct options *options); +static int run_test(iotype iot, parameters parms); +static void output_all_info(minmax *mm, int count, int indent_level); static void get_minmax(minmax *mm, double val); static minmax accumulate_minmax_stuff(minmax *mm, long raw_size, int count); static int create_comm_world(int num_procs, int *doing_pio); static int destroy_comm_world(void); -static void output_report(FILE *output, const char *fmt, ...); -static void print_indent(register FILE *output, register int indent); +static void output_report(const char *fmt, ...); +static void print_indent(register int indent); static void usage(const char *prog); /* @@ -246,9 +248,10 @@ main(int argc, char **argv) { int ret; int exit_value = EXIT_SUCCESS; - FILE *output = stdout; struct options *opts = NULL; + output = stdout; + /* initialize MPI and get the maximum num of processors we started with */ MPI_Init(&argc, &argv); ret = MPI_Comm_size(MPI_COMM_WORLD, &comm_world_nprocs_g); @@ -295,7 +298,7 @@ main(int argc, char **argv) } } - run_test_loop(output, opts); + run_test_loop(opts); finish: MPI_Finalize(); @@ -324,7 +327,7 @@ finish: * Modifications: */ static void -run_test_loop(FILE *output, struct options *opts) +run_test_loop(struct options *opts) { parameters parms; long num_procs; @@ -363,7 +366,7 @@ run_test_loop(FILE *output, struct options *opts) /* only processes doing PIO will run the tests */ if (doing_pio){ - output_report(output, "Number of processors = %ld\n", parms.num_procs); + output_report("Number of processors = %ld\n", parms.num_procs); /* multiply the xfer buffer size by 2 for each loop iteration */ for (buf_size = opts->min_xfer_size; @@ -371,23 +374,22 @@ run_test_loop(FILE *output, struct options *opts) parms.buf_size = buf_size; parms.num_elmts = opts->file_size / (parms.num_dsets * sizeof(int)); - print_indent(output, 1); - output_report(output, "Transfer Buffer Size: %ld bytes, File size: %.2f MBs\n", + print_indent(1); + output_report("Transfer Buffer Size: %ld bytes, File size: %.2f MBs\n", buf_size, ((double)parms.num_dsets * parms.num_elmts * sizeof(int)) / ONE_MB); - print_indent(output, 1); - output_report(output, - " # of files: %ld, # of dsets: %ld, # of elmts per dset: %ld\n", + print_indent(1); + output_report(" # of files: %ld, # of dsets: %ld, # of elmts per dset: %ld\n", parms.num_files, parms.num_dsets, parms.num_elmts); if (io_runs & PIO_RAW) - run_test(output, RAWIO, parms); + run_test(RAWIO, parms); if (io_runs & PIO_MPI) - run_test(output, MPIO, parms); + run_test(MPIO, parms); if (io_runs & PIO_HDF5) - run_test(output, PHDF5, parms); + run_test(PHDF5, parms); } if (destroy_comm_world() != SUCCESS) { @@ -405,7 +407,7 @@ run_test_loop(FILE *output, struct options *opts) * Modifications: */ static int -run_test(FILE *output, iotype iot, parameters parms) +run_test(iotype iot, parameters parms) { results res; register int i, ret_value = SUCCESS; @@ -423,18 +425,18 @@ run_test(FILE *output, iotype iot, parameters parms) raw_size = parms.num_dsets * parms.num_elmts * sizeof(int); parms.io_type = iot; - print_indent(output, 2); - output_report(output, "Type of IO = "); + print_indent(2); + output_report("Type of IO = "); switch (iot) { case RAWIO: - output_report(output, "Raw\n"); + output_report("Raw\n"); break; case MPIO: - output_report(output, "MPIO\n"); + output_report("MPIO\n"); break; case PHDF5: - output_report(output, "PHDF5\n"); + output_report("PHDF5\n"); break; } @@ -472,7 +474,8 @@ run_test(FILE *output, iotype iot, parameters parms) double t; MPI_Barrier(pio_comm_g); - res = do_pio(output, parms); + res = do_pio(parms); + timer = res.timers; /* gather all of the "write" times */ t = get_time(res.timers, HDF5_FINE_WRITE_FIXED_DIMS); @@ -497,90 +500,91 @@ run_test(FILE *output, iotype iot, parameters parms) get_minmax(&read_gross_mm, t); read_gross_mm_table[i] = read_gross_mm; + pio_time_destroy(res.timers); } /* accumulate and output the max, min, and average "write" times */ if (pio_debug_level >= 3) { /* output all of the times for all iterations */ - print_indent(output, 3); - output_report(output, "Write details:\n"); - output_all_info(output, write_mm_table, parms.num_iters, 4); + print_indent(3); + output_report("Write details:\n"); + output_all_info(write_mm_table, parms.num_iters, 4); } total_mm = accumulate_minmax_stuff(write_mm_table, raw_size, parms.num_iters); - print_indent(output, 3); - output_report(output, "Write (%d iteration(s)):\n", parms.num_iters); + print_indent(3); + output_report("Write (%d iteration(s)):\n", parms.num_iters); - print_indent(output, 4); - output_report(output, "Minimum Throughput: %.2f MB/s\n", total_mm.min); - print_indent(output, 4); - output_report(output, "Maximum Throughput: %.2f MB/s\n", total_mm.max); - print_indent(output, 4); - output_report(output, "Average Throughput: %.2f MB/s\n", + print_indent(4); + output_report("Minimum Throughput: %.2f MB/s\n", total_mm.min); + print_indent(4); + output_report("Maximum Throughput: %.2f MB/s\n", total_mm.max); + print_indent(4); + output_report("Average Throughput: %.2f MB/s\n", total_mm.sum / total_mm.num); /* accumulate and output the max, min, and average "gross write" times */ if (pio_debug_level >= 3) { /* output all of the times for all iterations */ - print_indent(output, 3); - output_report(output, "Write Open-Close details:\n"); - output_all_info(output, write_gross_mm_table, parms.num_iters, 4); + print_indent(3); + output_report("Write Open-Close details:\n"); + output_all_info(write_gross_mm_table, parms.num_iters, 4); } total_mm = accumulate_minmax_stuff(write_gross_mm_table, raw_size, parms.num_iters); - print_indent(output, 3); - output_report(output, "Write Open-Close (%d iteration(s)):\n", parms.num_iters); + print_indent(3); + output_report("Write Open-Close (%d iteration(s)):\n", parms.num_iters); - print_indent(output, 4); - output_report(output, "Minimum Throughput: %.2f MB/s\n", total_mm.min); - print_indent(output, 4); - output_report(output, "Maximum Throughput: %.2f MB/s\n", total_mm.max); - print_indent(output, 4); - output_report(output, "Average Throughput: %.2f MB/s\n", + print_indent(4); + output_report("Minimum Throughput: %.2f MB/s\n", total_mm.min); + print_indent(4); + output_report("Maximum Throughput: %.2f MB/s\n", total_mm.max); + print_indent(4); + output_report("Average Throughput: %.2f MB/s\n", total_mm.sum / total_mm.num); /* accumulate and output the max, min, and average "read" times */ if (pio_debug_level >= 3) { /* output all of the times for all iterations */ - print_indent(output, 3); - output_report(output, "Read details:\n"); - output_all_info(output, read_mm_table, parms.num_iters, 4); + print_indent(3); + output_report("Read details:\n"); + output_all_info(read_mm_table, parms.num_iters, 4); } total_mm = accumulate_minmax_stuff(read_mm_table, raw_size, parms.num_iters); - print_indent(output, 3); - output_report(output, "Read (%d iteration(s)):\n", parms.num_iters); + print_indent(3); + output_report("Read (%d iteration(s)):\n", parms.num_iters); - print_indent(output, 4); - output_report(output, "Minimum Throughput: %.2f MB/s\n", total_mm.min); - print_indent(output, 4); - output_report(output, "Maximum Throughput: %.2f MB/s\n", total_mm.max); - print_indent(output, 4); - output_report(output, "Average Throughput: %.2f MB/s\n", + print_indent(4); + output_report("Minimum Throughput: %.2f MB/s\n", total_mm.min); + print_indent(4); + output_report("Maximum Throughput: %.2f MB/s\n", total_mm.max); + print_indent(4); + output_report("Average Throughput: %.2f MB/s\n", total_mm.sum / total_mm.num); /* accumulate and output the max, min, and average "gross read" times */ if (pio_debug_level >= 3) { /* output all of the times for all iterations */ - print_indent(output, 3); - output_report(output, "Read Open-Close details:\n"); - output_all_info(output, read_gross_mm_table, parms.num_iters, 4); + print_indent(3); + output_report("Read Open-Close details:\n"); + output_all_info(read_gross_mm_table, parms.num_iters, 4); } total_mm = accumulate_minmax_stuff(read_gross_mm_table, raw_size, parms.num_iters); - print_indent(output, 3); - output_report(output, "Read Open-Close (%d iteration(s)):\n", parms.num_iters); + print_indent(3); + output_report("Read Open-Close (%d iteration(s)):\n", parms.num_iters); - print_indent(output, 4); - output_report(output, "Minimum Throughput: %.2f MB/s\n", total_mm.min); - print_indent(output, 4); - output_report(output, "Maximum Throughput: %.2f MB/s\n", total_mm.max); - print_indent(output, 4); - output_report(output, "Average Throughput: %.2f MB/s\n", + print_indent(4); + output_report("Minimum Throughput: %.2f MB/s\n", total_mm.min); + print_indent(4); + output_report("Maximum Throughput: %.2f MB/s\n", total_mm.max); + print_indent(4); + output_report("Average Throughput: %.2f MB/s\n", total_mm.sum / total_mm.num); /* clean up our mess */ @@ -588,7 +592,6 @@ run_test(FILE *output, iotype iot, parameters parms) free(read_mm_table); free(write_gross_mm_table); free(read_gross_mm_table); - pio_time_destroy(res.timers); return ret_value; } @@ -600,17 +603,17 @@ run_test(FILE *output, iotype iot, parameters parms) * Modifications: */ static void -output_all_info(FILE *output, minmax *mm, int count, int indent_level) +output_all_info(minmax *mm, int count, int indent_level) { register int i; for (i = 0; i < count; ++i) { - print_indent(output, indent_level); - output_report(output, "Iteration %d:\n", i + 1); - print_indent(output, indent_level + 1); - output_report(output, "Minimum Time: %.2fs\n", mm[i].min); - print_indent(output, indent_level + 1); - output_report(output, "Maximum Time: %.2fs\n", mm[i].max); + print_indent(indent_level); + output_report("Iteration %d:\n", i + 1); + print_indent(indent_level + 1); + output_report("Minimum Time: %.2fs\n", mm[i].min); + print_indent(indent_level + 1); + output_report("Maximum Time: %.2fs\n", mm[i].max); } } @@ -756,7 +759,7 @@ destroy_comm_world(void) * Modifications: */ static void -output_report(FILE *output, const char *fmt, ...) +output_report(const char *fmt, ...) { int myrank; @@ -780,7 +783,7 @@ output_report(FILE *output, const char *fmt, ...) * Modifications: */ static void -print_indent(register FILE *output, register int indent) +print_indent(register int indent) { int myrank; diff --git a/perform/pio_perf.h b/perform/pio_perf.h index 67beb3c..89cba72 100644 --- a/perform/pio_perf.h +++ b/perform/pio_perf.h @@ -40,8 +40,10 @@ typedef struct results_ { #define FAIL -1 #endif /* !FAIL */ -extern int comm_world_rank_g; /* my rank in MPI_COMM_RANK */ -extern int comm_world_nprocs_g;/* num. of processes of MPI_COMM_WORLD */ +extern FILE *output; /* output file */ +extern pio_time *timer; /* timer: global for stub functions */ +extern int comm_world_rank_g; /* my rank in MPI_COMM_RANK */ +extern int comm_world_nprocs_g;/* num. of processes of MPI_COMM_WORLD */ extern MPI_Comm pio_comm_g; /* Communicator to run the PIO */ extern int pio_mpi_rank_g; /* MPI rank of pio_comm_g */ extern int pio_mpi_nprocs_g; /* number of processes of pio_comm_g */ @@ -50,13 +52,14 @@ extern int pio_debug_level; /* The debug level: * 1 - Minimal * 2 - Some more * 3 - Maximal + * 4 - Even More Debugging (timer stuff) */ #ifdef __cplusplus extern "C" { #endif /* __cplusplus */ -extern results do_pio(FILE * output, parameters param); +extern results do_pio(parameters param); #ifdef __cplusplus } diff --git a/perform/pio_timer.c b/perform/pio_timer.c index 2ebdbce..1bf1b0c 100644 --- a/perform/pio_timer.c +++ b/perform/pio_timer.c @@ -19,6 +19,8 @@ #include <mpi.h> +#include "pio_perf.h" + /* * The number to divide the tv_usec field with to get a nice decimal to add to * the number of seconds. @@ -124,6 +126,47 @@ set_time(pio_time *pt, timer_type t, int start_stop) } } + if (pio_debug_level >= 4) { + char *msg; + int myrank; + + MPI_Comm_rank(pio_comm_g, &myrank); + + switch (t) { + case HDF5_FILE_OPENCLOSE: + msg = "File Open/Close"; + break; + case HDF5_DATASET_CREATE: + msg = "Dataset Create"; + break; + case HDF5_MPI_WRITE: + msg = "MPI Write"; + break; + case HDF5_MPI_READ: + msg = "MPI Read"; + break; + case HDF5_FINE_WRITE_FIXED_DIMS: + msg = "Fine Write"; + break; + case HDF5_FINE_READ_FIXED_DIMS: + msg = "Fine Read"; + break; + case HDF5_GROSS_WRITE_FIXED_DIMS: + msg = "Gross Write"; + break; + case HDF5_GROSS_READ_FIXED_DIMS: + msg = "Gross Read"; + break; + default: + msg = "Unknown Timer"; + break; + } + + fprintf(output, " Proc %d: %s %s: %.2f\n", myrank, msg, + (start_stop == START ? "Start" : "Stop"), + pt->total_time[t]); + } + return pt; } diff --git a/perform/pio_timer.h b/perform/pio_timer.h index dc723a0..4144825 100644 --- a/perform/pio_timer.h +++ b/perform/pio_timer.h @@ -22,6 +22,8 @@ typedef enum timer_type_ { HDF5_FILE_OPENCLOSE, HDF5_DATASET_CREATE, + HDF5_MPI_WRITE, + HDF5_MPI_READ, HDF5_FINE_WRITE_FIXED_DIMS, HDF5_FINE_READ_FIXED_DIMS, HDF5_GROSS_WRITE_FIXED_DIMS, |